Jana Diesner

Zhiwen You, Nafiseh Nikeghbal, Jana Diesner

Language models (LMs) can produce gendered language and stereotypes even when given neutral prompts. Most prior work on gender bias in LMs primarily examines gender through a binary lens (feminine vs. masculine), with limited attention to gender-neutral forms, such as they/them pronouns or neutrally phrased job titles. How gender-related signals are encoded in the internal representations of LMs remains an open question. In this work, we study gender-specific neurons in LMs across three categories: feminine, masculine, and gender-neutral. We propose a neuron-level intervention method to identify neurons that are strongly tied to each gender category. We then test these neurons through controlled generation, showing that activating or masking gender-related neurons can steer a sentence toward a target gender form while preserving its original meaning. To evaluate the effectiveness of our gender-intervention approach, we curate two datasets with controlled sentences labeled across all three gender categories and validate the data quality through human evaluation. Experiments on two open-source LMs show that gender-specific neurons are not evenly distributed across model layers; instead, they concentrate heavily in the earliest layers with smaller contributions from later layers. Compared to existing methods, our method achieves more precise gender control, with less leakage into non-target gender categories and stable output quality through two evaluation criteria. Overall, our work examines how gender is encoded in LMs and provides a simple yet effective approach toward controlled gender intervention for both neuron intervention evaluation and gender bias mitigation. Code and datasets are available at: https://github.com/zhiwenyou103/Gender-Neuron-Intervention

Yubin Ge, Ziang Xiao, Jana Diesner et al. · microsoft-research

Generating follow-up questions on the fly could significantly improve conversational survey quality and user experiences by enabling a more dynamic and personalized survey structure. In this paper, we proposed a novel task for knowledge-driven follow-up question generation in conversational surveys. We constructed a new human-annotated dataset of human-written follow-up questions with dialogue history and labeled knowledge in the context of conversational surveys. Along with the dataset, we designed and validated a set of reference-free Gricean-inspired evaluation metrics to systematically evaluate the quality of generated follow-up questions. We then propose a two-staged knowledge-driven model for the task, which generates informative and coherent follow-up questions by using knowledge to steer the generation process. The experiments demonstrate that compared to GPT-based baseline models, our two-staged model generates more informative, coherent, and clear follow-up questions.

Shubhanshu Mishra, Jana Diesner

Online data streams make training machine learning models hard because of distribution shift and new patterns emerging over time. For natural language processing (NLP) tasks that utilize a collection of features based on lexicons and rules, it is important to adapt these features to the changing data. To address this challenge we introduce PyTAIL, a python library, which allows a human in the loop approach to actively train NLP models. PyTAIL enhances generic active learning, which only suggests new instances to label by also suggesting new features like rules and lexicons to label. Furthermore, PyTAIL is flexible enough for users to accept, reject, or update rules and lexicons as the model is being trained. Finally, we simulate the performance of PyTAIL on existing social media benchmark datasets for text classification. We compare various active learning strategies on these benchmarks. The model closes the gap with as few as 10% of the training data. Finally, we also highlight the importance of tracking evaluation metric on remaining data (which is not yet merged with active learning) alongside the test dataset. This highlights the effectiveness of the model in accurately annotating the remaining dataset, which is especially suitable for batch processing of large unlabelled corpora. PyTAIL will be available at https://github.com/socialmediaie/pytail.

Amir Hossein Kargaran, Nafiseh Nikeghbal, Jana Diesner et al.

Optical character recognition (OCR) has advanced rapidly with the rise of vision-language models, yet evaluation has remained concentrated on a small cluster of high- and mid-resource scripts. We introduce GlotOCR Bench, a comprehensive benchmark evaluating OCR generalization across 100+ Unicode scripts. Our benchmark comprises clean and degraded image variants rendered from real multilingual texts. Images are rendered using fonts from the Google Fonts repository, shaped with HarfBuzz and rasterized with FreeType, supporting both LTR and RTL scripts. Samples of rendered images were manually reviewed to verify correct rendering across all scripts. We evaluate a broad suite of open-weight and proprietary vision-language models and find that most perform well on fewer than ten scripts, and even the strongest frontier models fail to generalize beyond thirty scripts. Performance broadly tracks script-level pretraining coverage, suggesting that current OCR systems rely on language model pretraining as much as on visual recognition. Models confronted with unfamiliar scripts either produce random noise or hallucinate characters from similar scripts they already know. We release the benchmark and pipeline for reproducibility. Pipeline Code: https://github.com/cisnlp/glotocr-bench, Benchmark: https://hf.co/datasets/cis-lmu/glotocr-bench.

Sullam Jeoung, Jana Diesner, Halil Kilicoglu

As language models continue to be integrated into applications of personal and societal relevance, ensuring these models' trustworthiness is crucial, particularly with respect to producing consistent outputs regardless of sensitive attributes. Given that first names may serve as proxies for (intersectional) socio-demographic representations, it is imperative to examine the impact of first names on commonsense reasoning capabilities. In this paper, we study whether a model's reasoning given a specific input differs based on the first names provided. Our underlying assumption is that the reasoning about Alice should not differ from the reasoning about James. We propose and implement a controlled experimental framework to measure the causal effect of first names on commonsense reasoning, enabling us to distinguish between model predictions due to chance and caused by actual factors of interest. Our results indicate that the frequency of first names has a direct effect on model prediction, with less frequent names yielding divergent predictions compared to more frequent names. To gain insights into the internal mechanisms of models that are contributing to these behaviors, we also conduct an in-depth explainable analysis. Overall, our findings suggest that to ensure model robustness, it is essential to augment datasets with more diverse first names during the configuration stage.

Sullam Jeoung, Jana Diesner

Previous work has examined how debiasing language models affect downstream tasks, specifically, how debiasing techniques influence task performance and whether debiased models also make impartial predictions in downstream tasks or not. However, what we don't understand well yet is why debiasing methods have varying impacts on downstream tasks and how debiasing techniques affect internal components of language models, i.e., neurons, layers, and attentions. In this paper, we decompose the internal mechanisms of debiasing language models with respect to gender by applying causal mediation analysis to understand the influence of debiasing methods on toxicity detection as a downstream task. Our findings suggest a need to test the effectiveness of debiasing methods with different bias metrics, and to focus on changes in the behavior of certain components of the models, e.g.,first two layers of language models, and attention heads.

Zhiwen You, HaeJin Lee, Shubhanshu Mishra et al.

Name-based gender prediction has traditionally categorized individuals as either female or male based on their names, using a binary classification system. That binary approach can be problematic in the cases of gender-neutral names that do not align with any one gender, among other reasons. Relying solely on binary gender categories without recognizing gender-neutral names can reduce the inclusiveness of gender prediction tasks. We introduce an additional gender category, i.e., "neutral", to study and address potential gender biases in Large Language Models (LLMs). We evaluate the performance of several foundational and large language models in predicting gender based on first names only. Additionally, we investigate the impact of adding birth years to enhance the accuracy of gender prediction, accounting for shifting associations between names and genders over time. Our findings indicate that most LLMs identify male and female names with high accuracy (over 80%) but struggle with gender-neutral names (under 40%), and the accuracy of gender prediction is higher for English-based first names than non-English names. The experimental results show that incorporating the birth year does not improve the overall accuracy of gender prediction, especially for names with evolving gender associations. We recommend using caution when applying LLMs for gender identification in downstream tasks, particularly when dealing with non-binary gender labels.

Sullam Jeoung, Yubin Ge, Jana Diesner

Large Language Models (LLMs) have been observed to encode and perpetuate harmful associations present in the training data. We propose a theoretically grounded framework called StereoMap to gain insights into their perceptions of how demographic groups have been viewed by society. The framework is grounded in the Stereotype Content Model (SCM); a well-established theory from psychology. According to SCM, stereotypes are not all alike. Instead, the dimensions of Warmth and Competence serve as the factors that delineate the nature of stereotypes. Based on the SCM theory, StereoMap maps LLMs' perceptions of social groups (defined by socio-demographic features) using the dimensions of Warmth and Competence. Furthermore, the framework enables the investigation of keywords and verbalizations of reasoning of LLMs' judgments to uncover underlying factors influencing their perceptions. Our results show that LLMs exhibit a diverse range of perceptions towards these groups, characterized by mixed evaluations along the dimensions of Warmth and Competence. Furthermore, analyzing the reasonings of LLMs, our findings indicate that LLMs demonstrate an awareness of social disparities, often stating statistical data and research findings to support their reasoning. This study contributes to the understanding of how LLMs perceive and represent social groups, shedding light on their potential biases and the perpetuation of harmful associations.

Shaghayegh Kolli, Richard Rosenbaum, Timo Cavelius et al.

Large language models (LLMs) excel in generating fluent utterances but can lack reliable grounding in verified information. At the same time, knowledge-graph-based fact-checkers deliver precise and interpretable evidence, yet suffer from limited coverage or latency. By integrating LLMs with knowledge graphs and real-time search agents, we introduce a hybrid fact-checking approach that leverages the individual strengths of each component. Our system comprises three autonomous steps: 1) a Knowledge Graph (KG) Retrieval for rapid one-hop lookups in DBpedia, 2) an LM-based classification guided by a task-specific labeling prompt, producing outputs with internal rule-based logic, and 3) a Web Search Agent invoked only when KG coverage is insufficient. Our pipeline achieves an F1 score of 0.93 on the FEVER benchmark on the Supported/Refuted split without task-specific fine-tuning. To address Not enough information cases, we conduct a targeted reannotation study showing that our approach frequently uncovers valid evidence for claims originally labeled as Not Enough Information (NEI), as confirmed by both expert annotators and LLM reviewers. With this paper, we present a modular, opensource fact-checking pipeline with fallback strategies and generalization across datasets.

Nafiseh Nikeghbal, Amir Hossein Kargaran, Jana Diesner

Improvements in model construction, including fortified safety guardrails, allow Large language models (LLMs) to increasingly pass standard safety checks. However, LLMs sometimes slip into revealing harmful behavior, such as expressing racist viewpoints, during conversations. To analyze this systematically, we introduce CoBia, a suite of lightweight adversarial attacks that allow us to refine the scope of conditions under which LLMs depart from normative or ethical behavior in conversations. CoBia creates a constructed conversation where the model utters a biased claim about a social group. We then evaluate whether the model can recover from the fabricated bias claim and reject biased follow-up questions. We evaluate 11 open-source as well as proprietary LLMs for their outputs related to six socio-demographic categories that are relevant to individual safety and fair treatment, i.e., gender, race, religion, nationality, sex orientation, and others. Our evaluation is based on established LLM-based bias metrics, and we compare the results against human judgments to scope out the LLMs' reliability and alignment. The results suggest that purposefully constructed conversations reliably reveal bias amplification and that LLMs often fail to reject biased follow-up questions during dialogue. This form of stress-testing highlights deeply embedded biases that can be surfaced through interaction. Code and artifacts are available at https://github.com/nafisenik/CoBia.

Pingjing Yang, Jennifer Cromley, Jana Diesner

Understanding how novices acquire and hone visual search skills is crucial for developing and optimizing training methods across domains. Network analysis methods can be used to analyze graph representations of visual expertise. This study investigates the relationship between eye-gaze movements and learning outcomes among undergraduate dentistry students who were diagnosing dental radiographs over multiple semesters. We use network analysis techniques to model eye-gaze scanpaths as directed graphs and examine changes in network metrics over time. Using time series clustering on each metric, we identify distinct patterns of visual search strategies and explore their association with students' diagnostic performance. Our findings suggest that the network metric of transition entropy is negatively correlated with performance scores, while the number of nodes and edges as well as average PageRank are positively correlated with performance scores. Changes in network metrics for individual students over time suggest a developmental shift from intermediate to expert-level processing. These insights contribute to understanding expertise acquisition in visual tasks and can inform the design of AI-assisted learning interventions.

HaeJin Lee, Shubhanshu Mishra, Apratim Mishra et al.

Gender biases in scholarly metrics remain a persistent concern, despite numerous bibliometric studies exploring their presence and absence across productivity, impact, acknowledgment, and self-citations. However, methodological inconsistencies, particularly in author name disambiguation and gender identification, limit the reliability and comparability of these studies, potentially perpetuating misperceptions and hindering effective interventions. A review of 70 relevant publications over the past 12 years reveals a wide range of approaches, from name-based and manual searches to more algorithmic and gold-standard methods, with no clear consensus on best practices. This variability, compounded by challenges such as accurately disambiguating Asian names and managing unassigned gender labels, underscores the urgent need for standardized and robust methodologies. To address this critical gap, we propose the development and implementation of ``Scholarly Data Analysis (SoDA) Cards." These cards will provide a structured framework for documenting and reporting key methodological choices in scholarly data analysis, including author name disambiguation and gender identification procedures. By promoting transparency and reproducibility, SoDA Cards will facilitate more accurate comparisons and aggregations of research findings, ultimately supporting evidence-informed policymaking and enabling the longitudinal tracking of analytical approaches in the study of gender and other social biases in academia.

Sullam Jeoung, Yubin Ge, Haohan Wang et al.

Examining the alignment of large language models (LLMs) has become increasingly important, e.g., when LLMs fail to operate as intended. This study examines the alignment of LLMs with human values for the domain of politics. Prior research has shown that LLM-generated outputs can include political leanings and mimic the stances of political parties on various issues. However, the extent and conditions under which LLMs deviate from empirical positions are insufficiently examined. To address this gap, we analyze the factors that contribute to LLMs' deviations from empirical positions on political issues, aiming to quantify these deviations and identify the conditions that cause them. Drawing on findings from cognitive science about representativeness heuristics, i.e., situations where humans lean on representative attributes of a target group in a way that leads to exaggerated beliefs, we scrutinize LLM responses through this heuristics' lens. We conduct experiments to determine how LLMs inflate predictions about political parties, which results in stereotyping. We find that while LLMs can mimic certain political parties' positions, they often exaggerate these positions more than human survey respondents do. Also, LLMs tend to overemphasize representativeness more than humans. This study highlights the susceptibility of LLMs to representativeness heuristics, suggesting a potential vulnerability of LLMs that facilitates political stereotyping. We also test prompt-based mitigation strategies, finding that strategies that can mitigate representative heuristics in humans are also effective in reducing the influence of representativeness on LLM-generated responses.

M. Janina Sarol, Ly Dinh, Rezvaneh Rezapour et al.

In times of crisis, identifying the essential needs is a crucial step to providing appropriate resources and services to affected entities. Social media platforms such as Twitter contain vast amount of information about the general public's needs. However, the sparsity of the information as well as the amount of noisy content present a challenge to practitioners to effectively identify shared information on these platforms. In this study, we propose two novel methods for two distinct but related needs detection tasks: the identification of 1) a list of resources needed ranked by priority, and 2) sentences that specify who-needs-what resources. We evaluated our methods on a set of tweets about the COVID-19 crisis. For task 1 (detecting top needs), we compared our results against two given lists of resources and achieved 64% precision. For task 2 (detecting who-needs-what), we compared our results on a set of 1,000 annotated tweets and achieved a 68% F1-score.

Ming Jiang, Junjie Hu, Qiuyuan Huang et al.

Popular metrics used for evaluating image captioning systems, such as BLEU and CIDEr, provide a single score to gauge the system's overall effectiveness. This score is often not informative enough to indicate what specific errors are made by a given system. In this study, we present a fine-grained evaluation method REO for automatically measuring the performance of image captioning systems. REO assesses the quality of captions from three perspectives: 1) Relevance to the ground truth, 2) Extraness of the content that is irrelevant to the ground truth, and 3) Omission of the elements in the images and human references. Experiments on three benchmark datasets demonstrate that our method achieves a higher consistency with human judgments and provides more intuitive evaluation results than alternative metrics.

Ming Jiang, Qiuyuan Huang, Lei Zhang et al.

This paper presents a new metric called TIGEr for the automatic evaluation of image captioning systems. Popular metrics, such as BLEU and CIDEr, are based solely on text matching between reference captions and machine-generated captions, potentially leading to biased evaluations because references may not fully cover the image content and natural language is inherently ambiguous. Building upon a machine-learned text-image grounding model, TIGEr allows to evaluate caption quality not only based on how well a caption represents image content, but also on how well machine-generated captions match human-generated captions. Our empirical tests show that TIGEr has a higher consistency with human judgments than alternative existing metrics. We also comprehensively assess the metric's effectiveness in caption evaluation by measuring the correlation between human judgments and metric scores.